Learning the Robust and Structured Control of Unknown Linear Systems

In many control systems in practice, the control input is subject to some specific structure and affected by noise. Designing control laws to stabilize these systems is challenging, especially when the system model is uncertain. This paper will solve this problem for linear systems with unknown system state matrix. First, a model-based framework is formulated to embed the structural control constraint and input noise into the linear quadratic regulator (LQR) setting to find the stabilizing control. Then, this model-based LQR setting is transformed into a databased learning framework, robust structured reinforcement learning (RSRL), to cope with the unknown system state matrix. As a result, the data-based control can stabilize the unknown system with input noise, while obeying the structural constraint. Theoretical results will be validated on a multi-agent network with 6 agents.